ERGO: Event Relational Graph Transformer for Document-level Event Causality Identification

Proceedings of the 29th International Conference on Computational Linguistics, pages 2118–2128 October 12–17, 2022. 2118 ERGO: Event Relational Graph Transformer for Document-level Event Causality Identification Meiqi Chen1, Yixin Cao2, Kunquan Deng3, Mukai Li4, Kun Wang4, Jing Shao4, Yan Zhang1∗ 1 Peking University 2 Singapore Management University 3 Beihang University 4 SenseTime Research meiqichen@stu.pku.edu.cn Abstract Document-level Event Causality Identification (DECI) aims to identify event-event causal re- lations in a document. Existing works usu- ally build an event graph for global reasoning across multiple sentences. However, the edges between events have to be carefully designed through heuristic rules or external tools. In this paper, we propose a novel Event Relational Graph TransfOrmer (ERGO) framework1 for DECI, to ease the graph construction and im- prove it over the noisy edge issue. Different from conventional event graphs, we define a pair of events as a node and build a complete event relational graph without any prior knowl- edge or tools. This naturally formulates DECI as a node classification problem, and thus we capture the causation transitivity among event pairs via a graph transformer. Furthermore, we design a criss-cross constraint and an adap- tive focal loss for the imbalanced classification, to alleviate the issues of false positives and false negatives. Extensive experiments on two benchmark datasets show that ERGO greatly outperforms previous state-of-the-art (SOTA) methods (12.8% F1 gains on average). 1 Introduction Event Causality Identification (ECI) is the task of identifying if the occurrence of one event causes another in text. As shown in Figure 1, given the text “... the outage2 was caused by a terrestrial break in the fiber in Egypt ...”, where “outage2” and “break” are event triggers, an ECI model should predict if they have a causal relation or not. Discovering causal relationships not only helps to deeply un- derstand how the world progresses, but also is an important goal of empirical research in various ar- eas, such as machine reading comprehension (Be- rant et al., 2014), question answering (Oh et al., 2016), future event forecasting (Hashimoto, 2019), ∗Corresponding author. 1https://github.com/chenmeiqii/ERGO.git Figure 1: Example of DECI. Solid purple lines denote target causal relations. and event knowledge graph construction (Ma et al., 2022b). Causality is usually implicit in natural lan- guage (Copley and Martin, 2014), especially when events scatter in a document, a.k.a. Document-level ECI (DECI). Recent methods typically construct an event graph to assist the global inference across multiple sentences, where nodes are events and edges are their relations, such as linguistic depen- dency or adjacent contexts (Gao et al., 2019; Zhao et al., 2021). However, there are two major issues. First, the edges heavily rely on external tools and heuristic rules, which are not always reliable and may introduce noise (Tran Phu and Nguyen, 2021). Second, it is like a chicken-egg problem — to iden- tify (causal) relations between events, you need to extract their relations to build the graph first. In this paper, we propose a novel Event Relational Graph TransfOrmer (ERGO) frame- work for DECI, which doesn’t require …

Content truncated to 500 words